Sound features gave an uncertain gain beyond signal history after five minutes of training in nine Podcast SEEG recordings with a common average reference

We tested whether sound features improve later SEEG high-gamma amplitude prediction beyond recent signal history. We used nine public Podcast recordings and all 325 eligible depth contacts. Models used the first five minutes for training and minutes 20 to 30 for testing. The mean gain was 0.193% of test signal variance. The 95% participant bootstrap interval was -0.220% to +0.604%. Six of nine participants had a positive gain. The planned model does not establish a positive mean gain. Ten-minute training gave a larger gain in a secondary test. Aligned sound performed better than eight fixed sound shifts in all nine participants. These checks are exploratory. Results apply to one story, a fixed ridge model, and released signals with a common average reference. Whole-record preprocessing limits interpretation to existing processed signals. The package preserves compact inputs, code, models, predictions, scores, source versions and hashes, public source checks, and reproduction instructions. A separate solver checked the fits. A full repeat gave identical result files.

SEEG · ECoG · high-gamma amplitude · naturalistic listening · sound features · signal history · ridge regression · out-of-sample prediction · Podcast dataset